Pytorch-lightning: Training accuracy

[lorenzoFabbri]

I was wondering whether there is something like validation_end but for training (e.g., training_end). I want to compute the training accuracy at the end of each epoch. Thanks!

[williamFalcon]

you can use hooks
(on_epoch_end)

https://williamfalcon.github.io/pytorch-lightning/Trainer/hooks/

there's a typo in that doc... but you're looking for on_epoch_end

[lorenzoFabbri]

Thanks but I'm having some difficulties.
If I save the accuracy from each training step, how can I compute the average accuracy? on_epoch_end does not take anything as input, so I cannot write something similar to validation_end. Thanks.

[expectopatronum]

@lorenzoFabbri Did you figure out how to do this? I am also trying to do this but don't know how...

[lorenzoFabbri]

Hi. I actually decided not to use this library for that project. So I cannot help you!

[expectopatronum]

@lorenzoFabbri Thanks a lot for the fast response!

[Yevgnen]

@lorenzoFabbri Do you have any suggestion for replacement?

[lorenzoFabbri]

@Yevgnen Hi. I haven't use this library in a long time! Perhaps thins have changed in the meanwhile so I cannot really help you. I plan on giving it another try in the next few weeks, though.

[Yevgnen]

@lorenzoFabbri Thanks for quick reply! May I ask which library you are using?😅

[williamFalcon]

we introduced a full callbacks system now! give it a shot.

you can calculate the accuracy on epoch_end.

we’re also about to introduce:
on_training_epoch_end.

also, FYI: we now support TPUs and have a built in performance profiler

[lorenzoFabbri]

@Yevgnen I actually used plain PyTorch. I didn't have to do fancy things, so I wrote my own functions for training and testing.